Finding Stable Extremal Region Boundaries

This paper introduces a novel edge detection framework finding the most stable region boundaries in grayscale images. In contrast to common detection algorithms as Canny, which only analyze local discontinuities in image brightness our method also integrates mid-level information by analyzing regions that support the local gradient magnitudes. We build a component tree where every node contains a single connected region obtained from thresholding the input image and edges define the spatial relations between the nodes. Then connected nodes at different levels in the tree are compared by a simplified chamfer matching method. Matching regions have boundaries that stay similar over several image intensities, and thus contribute as stable edges to the final result. Since the component tree can be calculated in linear time and chamfer matching between nodes in the component tree is reduced to an analysis of the the distance transformation, results are obtained in a very efficient
manner. Furthermore, the proposed detection algorithm automatically labels all identified edges during calculation avoiding the required post-processing of connecting and labeling edge responses for usage in a detection framework. We compare our method against standard Canny and the Berkeley edge detector on the ETHZ shape classes and the Weizmann horses datasets, which both demonstrate better performance in reducing clutter.